Implicit Depth Deep Unfolding for 3-D SAR Tomography

Abstract

The synthetic aperture radar (SAR) tomography (TomoSAR) plays a crucial role in 3-D scene reconstruction and has been widely applied in various fields, such as urban mapping and disaster monitoring. However, in technical practice, the traditional compressed sensing (CS)-based algorithms suffer from issues, such as high computational complexity and difficult parameter estimation. In recent years, with the widespread application of deep networks, deep unfolding (DU)-based algorithms have alleviated the shortcomings of CS-based algorithms to a certain extent, but they are still limited by memory bottlenecks, which restricts the expansion of network layers. To address these problems, this letter introduces the deep equilibrium (DEQ) framework into the field of TomoSAR imaging for the first time and proposes a novel deep learning method based on this framework. The proposed DEQ-based network breaks through the layer dependency limitation of DU networks (DUNs) by directly solving equilibrium equations and can achieve theoretically infinite network depth iteration without increasing memory requirements. The simulation and experimental results demonstrate that the proposed method significantly improves the memory efficiency of DUNs while maintaining high imaging accuracy.